1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LDLM
27 # include <linux/slab.h>
28 # include <linux/dcache.h>
29 # include <linux/namei.h>
30 # include <linux/module.h>
31 # include <linux/lustre_dlm.h>
33 # include <liblustre.h>
34 # include <libcfs/kp30.h>
37 #include <linux/obd_class.h>
38 #include "ldlm_internal.h"
40 //struct lustre_lock ldlm_everything_lock;
43 char *ldlm_lockname[] = {
53 char *ldlm_typename[] = {
60 char *ldlm_it2str(int it)
67 case (IT_OPEN | IT_CREAT):
82 CERROR("Unknown intent %d\n", it);
87 extern kmem_cache_t *ldlm_lock_slab;
88 struct lustre_lock ldlm_handle_lock;
90 static ldlm_processing_policy ldlm_processing_policy_table[] = {
91 [LDLM_PLAIN] ldlm_process_plain_lock,
92 [LDLM_EXTENT] ldlm_process_extent_lock,
94 [LDLM_FLOCK] ldlm_process_flock_lock,
96 [LDLM_IBITS] ldlm_process_inodebits_lock,
99 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
101 return ldlm_processing_policy_table[res->lr_type];
104 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
110 * REFCOUNTED LOCK OBJECTS
115 * Lock refcounts, during creation:
116 * - one special one for allocation, dec'd only once in destroy
117 * - one for being a lock that's in-use
118 * - one for the addref associated with a new lock
120 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
122 atomic_inc(&lock->l_refc);
126 void ldlm_lock_put(struct ldlm_lock *lock)
130 if (atomic_dec_and_test(&lock->l_refc)) {
131 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
133 l_lock(&ns->ns_lock);
134 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
135 LASSERT(lock->l_destroyed);
136 LASSERT(list_empty(&lock->l_res_link));
138 spin_lock(&ns->ns_counter_lock);
140 spin_unlock(&ns->ns_counter_lock);
142 ldlm_resource_putref(lock->l_resource);
143 lock->l_resource = NULL;
145 class_export_put(lock->l_export);
148 LDLM_LOCK_PUT(lock->l_parent);
150 if (lock->l_lvb_data != NULL)
151 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
153 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
154 l_unlock(&ns->ns_lock);
160 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
163 l_lock(&lock->l_resource->lr_namespace->ns_lock);
164 if (!list_empty(&lock->l_lru)) {
165 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
166 list_del_init(&lock->l_lru);
167 lock->l_resource->lr_namespace->ns_nr_unused--;
168 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
170 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
174 /* This used to have a 'strict' flact, which recovery would use to mark an
175 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
176 * shall explain why it's gone: with the new hash table scheme, once you call
177 * ldlm_lock_destroy, you can never drop your final references on this lock.
178 * Because it's not in the hash table anymore. -phil */
179 void ldlm_lock_destroy(struct ldlm_lock *lock)
182 l_lock(&lock->l_resource->lr_namespace->ns_lock);
184 if (!list_empty(&lock->l_children)) {
185 LDLM_ERROR(lock, "still has children (%p)!",
186 lock->l_children.next);
187 ldlm_lock_dump(D_ERROR, lock, 0);
190 if (lock->l_readers || lock->l_writers) {
191 LDLM_ERROR(lock, "lock still has references");
192 ldlm_lock_dump(D_ERROR, lock, 0);
196 if (!list_empty(&lock->l_res_link)) {
197 LDLM_ERROR(lock, "lock still on resource");
198 ldlm_lock_dump(D_ERROR, lock, 0);
202 if (lock->l_destroyed) {
203 LASSERT(list_empty(&lock->l_lru));
204 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
208 lock->l_destroyed = 1;
210 list_del_init(&lock->l_export_chain);
211 ldlm_lock_remove_from_lru(lock);
212 class_handle_unhash(&lock->l_handle);
215 /* Wake anyone waiting for this lock */
216 /* FIXME: I should probably add yet another flag, instead of using
217 * l_export to only call this on clients */
219 class_export_put(lock->l_export);
220 lock->l_export = NULL;
221 if (lock->l_export && lock->l_completion_ast)
222 lock->l_completion_ast(lock, 0);
225 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
230 /* this is called by portals_handle2object with the handle lock taken */
231 static void lock_handle_addref(void *lock)
233 LDLM_LOCK_GET((struct ldlm_lock *)lock);
237 * usage: pass in a resource on which you have done ldlm_resource_get
238 * pass in a parent lock on which you have done a ldlm_lock_get
239 * after return, ldlm_*_put the resource and parent
240 * returns: lock with refcount 2 - one for current caller and one for remote
242 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
243 struct ldlm_resource *resource)
245 struct ldlm_lock *lock;
248 if (resource == NULL)
251 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
255 lock->l_resource = ldlm_resource_getref(resource);
257 atomic_set(&lock->l_refc, 2);
258 INIT_LIST_HEAD(&lock->l_children);
259 INIT_LIST_HEAD(&lock->l_res_link);
260 INIT_LIST_HEAD(&lock->l_lru);
261 INIT_LIST_HEAD(&lock->l_export_chain);
262 INIT_LIST_HEAD(&lock->l_pending_chain);
263 init_waitqueue_head(&lock->l_waitq);
265 spin_lock(&resource->lr_namespace->ns_counter_lock);
266 resource->lr_namespace->ns_locks++;
267 spin_unlock(&resource->lr_namespace->ns_counter_lock);
269 if (parent != NULL) {
270 l_lock(&parent->l_resource->lr_namespace->ns_lock);
271 lock->l_parent = LDLM_LOCK_GET(parent);
272 list_add(&lock->l_childof, &parent->l_children);
273 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
276 INIT_LIST_HEAD(&lock->l_handle.h_link);
277 class_handle_hash(&lock->l_handle, lock_handle_addref);
282 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
283 struct ldlm_res_id new_resid)
285 struct ldlm_resource *oldres = lock->l_resource;
288 l_lock(&ns->ns_lock);
289 if (memcmp(&new_resid, &lock->l_resource->lr_name,
290 sizeof(lock->l_resource->lr_name)) == 0) {
292 l_unlock(&ns->ns_lock);
296 LASSERT(new_resid.name[0] != 0);
298 /* This function assumes that the lock isn't on any lists */
299 LASSERT(list_empty(&lock->l_res_link));
301 lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
302 lock->l_resource->lr_type, 1);
303 if (lock->l_resource == NULL) {
308 /* ...and the flowers are still standing! */
309 ldlm_resource_putref(oldres);
311 l_unlock(&ns->ns_lock);
319 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
321 lockh->cookie = lock->l_handle.h_cookie;
324 /* if flags: atomically get the lock and set the flags.
325 * Return NULL if flag already set
328 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
330 struct ldlm_namespace *ns;
331 struct ldlm_lock *lock = NULL, *retval = NULL;
336 lock = class_handle2object(handle->cookie);
340 LASSERT(lock->l_resource != NULL);
341 ns = lock->l_resource->lr_namespace;
344 l_lock(&ns->ns_lock);
346 /* It's unlikely but possible that someone marked the lock as
347 * destroyed after we did handle2object on it */
348 if (lock->l_destroyed) {
349 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
354 if (flags && (lock->l_flags & flags)) {
360 lock->l_flags |= flags;
365 l_unlock(&ns->ns_lock);
369 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
370 struct lustre_handle *handle)
372 struct ldlm_lock *retval = NULL;
374 l_lock(&ns->ns_lock);
375 retval = __ldlm_handle2lock(handle, 0);
376 l_unlock(&ns->ns_lock);
381 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
383 ldlm_res2desc(lock->l_resource, &desc->l_resource);
384 desc->l_req_mode = lock->l_req_mode;
385 desc->l_granted_mode = lock->l_granted_mode;
386 memcpy(&desc->l_policy_data, &lock->l_policy_data,
387 sizeof(desc->l_policy_data));
390 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
391 void *data, int datalen)
393 struct ldlm_ast_work *w;
396 l_lock(&lock->l_resource->lr_namespace->ns_lock);
397 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
400 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
402 OBD_ALLOC(w, sizeof(*w));
409 w->w_datalen = datalen;
411 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
412 lock->l_flags |= LDLM_FL_AST_SENT;
413 /* If the enqueuing client said so, tell the AST recipient to
414 * discard dirty data, rather than writing back. */
415 if (new->l_flags & LDLM_AST_DISCARD_DATA)
416 lock->l_flags |= LDLM_FL_DISCARD_DATA;
418 ldlm_lock2desc(new, &w->w_desc);
421 w->w_lock = LDLM_LOCK_GET(lock);
422 list_add(&w->w_list, lock->l_resource->lr_tmp);
425 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
428 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
430 struct ldlm_lock *lock;
432 lock = ldlm_handle2lock(lockh);
433 ldlm_lock_addref_internal(lock, mode);
437 /* only called for local locks */
438 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
440 l_lock(&lock->l_resource->lr_namespace->ns_lock);
441 ldlm_lock_remove_from_lru(lock);
442 if (mode & (LCK_NL | LCK_CR | LCK_PR))
444 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
446 lock->l_last_used = jiffies;
448 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
449 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
452 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
454 struct ldlm_namespace *ns;
457 ns = lock->l_resource->lr_namespace;
459 l_lock(&ns->ns_lock);
460 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
461 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
462 LASSERT(lock->l_readers > 0);
465 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
466 LASSERT(lock->l_writers > 0);
470 if (lock->l_flags & LDLM_FL_LOCAL &&
471 !lock->l_readers && !lock->l_writers) {
472 /* If this is a local lock on a server namespace and this was
473 * the last reference, cancel the lock. */
474 CDEBUG(D_INFO, "forcing cancel of local lock\n");
475 lock->l_flags |= LDLM_FL_CBPENDING;
478 if (!lock->l_readers && !lock->l_writers &&
479 (lock->l_flags & LDLM_FL_CBPENDING)) {
480 /* If we received a blocked AST and this was the last reference,
481 * run the callback. */
482 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
483 CERROR("FL_CBPENDING set on non-local lock--just a "
486 LDLM_DEBUG(lock, "final decref done on cbpending lock");
488 LDLM_LOCK_GET(lock); /* dropped by bl thread */
489 ldlm_lock_remove_from_lru(lock);
490 l_unlock(&ns->ns_lock);
491 if (ldlm_bl_to_thread(ns, NULL, lock) != 0)
492 ldlm_handle_bl_callback(ns, NULL, lock);
493 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
494 !lock->l_readers && !lock->l_writers) {
495 /* If this is a client-side namespace and this was the last
496 * reference, put it on the LRU. */
497 LASSERT(list_empty(&lock->l_lru));
498 LASSERT(ns->ns_nr_unused >= 0);
499 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
501 l_unlock(&ns->ns_lock);
502 ldlm_cancel_lru(ns, LDLM_ASYNC);
504 l_unlock(&ns->ns_lock);
507 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
512 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
514 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
515 LASSERT(lock != NULL);
516 ldlm_lock_decref_internal(lock, mode);
520 /* This will drop a lock reference and mark it for destruction, but will not
521 * necessarily cancel the lock before returning. */
522 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
524 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
527 LASSERT(lock != NULL);
529 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
530 l_lock(&lock->l_resource->lr_namespace->ns_lock);
531 lock->l_flags |= LDLM_FL_CBPENDING;
532 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
533 ldlm_lock_decref_internal(lock, mode);
538 * - ldlm_lock_enqueue
539 * - ldlm_reprocess_queue
540 * - ldlm_lock_convert
542 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
545 struct ldlm_resource *res = lock->l_resource;
548 l_lock(&lock->l_resource->lr_namespace->ns_lock);
549 lock->l_granted_mode = lock->l_req_mode;
550 ldlm_resource_add_lock(res, &res->lr_granted, lock);
552 if (lock->l_granted_mode < res->lr_most_restr)
553 res->lr_most_restr = lock->l_granted_mode;
555 if (run_ast && lock->l_completion_ast != NULL)
556 ldlm_add_ast_work_item(lock, NULL, data, datalen);
558 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
562 /* returns a referenced lock or NULL. See the flag descriptions below, in the
563 * comment above ldlm_lock_match */
564 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
565 ldlm_policy_data_t *policy,
566 struct ldlm_lock *old_lock, int flags)
568 struct ldlm_lock *lock;
569 struct list_head *tmp;
571 list_for_each(tmp, queue) {
572 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
574 if (lock == old_lock)
577 /* llite sometimes wants to match locks that will be
578 * canceled when their users drop, but we allow it to match
579 * if it passes in CBPENDING and the lock still has users.
580 * this is generally only going to be used by children
581 * whose parents already hold a lock so forward progress
582 * can still happen. */
583 if (lock->l_flags & LDLM_FL_CBPENDING &&
584 !(flags & LDLM_FL_CBPENDING))
586 if (lock->l_flags & LDLM_FL_CBPENDING &&
587 lock->l_readers == 0 && lock->l_writers == 0)
590 if (!(lock->l_req_mode & mode))
593 if (lock->l_resource->lr_type == LDLM_EXTENT &&
594 (lock->l_policy_data.l_extent.start >
595 policy->l_extent.start ||
596 lock->l_policy_data.l_extent.end < policy->l_extent.end))
599 if (lock->l_resource->lr_type == LDLM_EXTENT &&
601 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
604 /* We match if we have existing lock with same or wider set
606 if (lock->l_resource->lr_type == LDLM_IBITS &&
607 ((lock->l_policy_data.l_inodebits.bits &
608 policy->l_inodebits.bits) !=
609 policy->l_inodebits.bits))
612 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
615 if ((flags & LDLM_FL_LOCAL_ONLY) &&
616 !(lock->l_flags & LDLM_FL_LOCAL))
619 if (flags & LDLM_FL_TEST_LOCK)
622 ldlm_lock_addref_internal(lock, mode);
629 void ldlm_lock_allow_match(struct ldlm_lock *lock)
631 l_lock(&lock->l_resource->lr_namespace->ns_lock);
632 lock->l_flags |= LDLM_FL_CAN_MATCH;
633 wake_up(&lock->l_waitq);
634 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
637 /* Can be called in two ways:
639 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
640 * for a duplicate of.
642 * Otherwise, all of the fields must be filled in, to match against.
644 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
645 * server (ie, connh is NULL)
646 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
647 * list will be considered
648 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
649 * to be canceled can still be matched as long as they still have reader
650 * or writer refernces
651 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
652 * just tell us if we would have matched.
654 * Returns 1 if it finds an already-existing lock that is compatible; in this
655 * case, lockh is filled in with a addref()ed lock
657 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
658 struct ldlm_res_id *res_id, __u32 type,
659 ldlm_policy_data_t *policy, ldlm_mode_t mode,
660 struct lustre_handle *lockh)
662 struct ldlm_resource *res;
663 struct ldlm_lock *lock, *old_lock = NULL;
668 old_lock = ldlm_handle2lock(lockh);
671 ns = old_lock->l_resource->lr_namespace;
672 res_id = &old_lock->l_resource->lr_name;
673 type = old_lock->l_resource->lr_type;
674 mode = old_lock->l_req_mode;
677 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
679 LASSERT(old_lock == NULL);
683 l_lock(&ns->ns_lock);
685 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
688 if (flags & LDLM_FL_BLOCK_GRANTED)
690 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
693 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
699 ldlm_resource_putref(res);
700 l_unlock(&ns->ns_lock);
703 ldlm_lock2handle(lock, lockh);
704 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
705 struct l_wait_info lwi;
706 if (lock->l_completion_ast) {
707 int err = lock->l_completion_ast(lock,
708 LDLM_FL_WAIT_NOREPROC,
716 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
718 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
719 l_wait_event(lock->l_waitq,
720 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
726 l_lock(&ns->ns_lock);
727 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
728 type == LDLM_PLAIN ? res_id->name[2] :
729 policy->l_extent.start,
730 type == LDLM_PLAIN ? res_id->name[3] :
731 policy->l_extent.end);
732 l_unlock(&ns->ns_lock);
733 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
734 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
735 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
736 type, mode, res_id->name[0], res_id->name[1],
737 type == LDLM_PLAIN ? res_id->name[2] :
738 policy->l_extent.start,
739 type == LDLM_PLAIN ? res_id->name[3] :
740 policy->l_extent.end);
744 LDLM_LOCK_PUT(old_lock);
745 if (flags & LDLM_FL_TEST_LOCK && rc)
751 /* Returns a referenced lock */
752 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
753 struct lustre_handle *parent_lock_handle,
754 struct ldlm_res_id res_id, __u32 type,
756 ldlm_blocking_callback blocking,
757 ldlm_completion_callback completion,
758 ldlm_glimpse_callback glimpse,
759 void *data, __u32 lvb_len)
761 struct ldlm_resource *res, *parent_res = NULL;
762 struct ldlm_lock *lock, *parent_lock = NULL;
765 if (parent_lock_handle) {
766 parent_lock = ldlm_handle2lock(parent_lock_handle);
768 parent_res = parent_lock->l_resource;
771 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
775 lock = ldlm_lock_new(parent_lock, res);
776 ldlm_resource_putref(res);
777 if (parent_lock != NULL)
778 LDLM_LOCK_PUT(parent_lock);
783 lock->l_req_mode = mode;
784 lock->l_ast_data = data;
785 lock->l_blocking_ast = blocking;
786 lock->l_completion_ast = completion;
787 lock->l_glimpse_ast = glimpse;
788 lock->l_pid = current->pid;
791 lock->l_lvb_len = lvb_len;
792 OBD_ALLOC(lock->l_lvb_data, lvb_len);
793 if (lock->l_lvb_data == NULL) {
794 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
802 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
803 struct ldlm_lock **lockp,
804 void *cookie, int *flags)
806 struct ldlm_lock *lock = *lockp;
807 struct ldlm_resource *res = lock->l_resource;
808 int local = res->lr_namespace->ns_client;
809 ldlm_processing_policy policy;
810 ldlm_error_t rc = ELDLM_OK;
813 /* policies are not executed on the client or during replay */
814 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
815 && !local && ns->ns_policy) {
816 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
818 if (rc == ELDLM_LOCK_REPLACED) {
819 /* The lock that was returned has already been granted,
820 * and placed into lockp. If it's not the same as the
821 * one we passed in, then destroy the old one and our
822 * work here is done. */
823 if (lock != *lockp) {
824 ldlm_lock_destroy(lock);
827 *flags |= LDLM_FL_LOCK_CHANGED;
829 } else if (rc == ELDLM_LOCK_ABORTED ||
830 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
831 ldlm_lock_destroy(lock);
834 LASSERT(rc == ELDLM_OK);
837 l_lock(&ns->ns_lock);
838 if (local && lock->l_req_mode == lock->l_granted_mode) {
839 /* The server returned a blocked lock, but it was granted before
840 * we got a chance to actually enqueue it. We don't need to do
842 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
843 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
847 /* Some flags from the enqueue want to make it into the AST, via the
849 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
851 /* This distinction between local lock trees is very important; a client
852 * namespace only has information about locks taken by that client, and
853 * thus doesn't have enough information to decide for itself if it can
854 * be granted (below). In this case, we do exactly what the server
855 * tells us to do, as dictated by the 'flags'.
857 * We do exactly the same thing during recovery, when the server is
858 * more or less trusting the clients not to lie.
860 * FIXME (bug 268): Detect obvious lies by checking compatibility in
861 * granted/converting queues. */
862 ldlm_resource_unlink_lock(lock);
864 if (*flags & LDLM_FL_BLOCK_CONV)
865 ldlm_resource_add_lock(res, &res->lr_converting, lock);
866 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
867 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
869 ldlm_grant_lock(lock, NULL, 0, 0);
871 } else if (*flags & LDLM_FL_REPLAY) {
872 if (*flags & LDLM_FL_BLOCK_CONV) {
873 ldlm_resource_add_lock(res, &res->lr_converting, lock);
875 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
876 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
878 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
879 ldlm_grant_lock(lock, NULL, 0, 0);
882 /* If no flags, fall through to normal enqueue path. */
885 policy = ldlm_processing_policy_table[res->lr_type];
886 policy(lock, flags, 1, &rc);
889 l_unlock(&ns->ns_lock);
893 /* Must be called with namespace taken: queue is waiting or converting. */
894 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
896 struct list_head *tmp, *pos;
897 ldlm_processing_policy policy;
899 int rc = LDLM_ITER_CONTINUE;
903 policy = ldlm_processing_policy_table[res->lr_type];
906 list_for_each_safe(tmp, pos, queue) {
907 struct ldlm_lock *pending;
908 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
910 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
913 rc = policy(pending, &flags, 0, &err);
914 if (rc != LDLM_ITER_CONTINUE)
921 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
923 struct list_head *tmp, *pos;
927 l_check_no_ns_lock(ns);
929 list_for_each_safe(tmp, pos, rpc_list) {
930 struct ldlm_ast_work *w =
931 list_entry(tmp, struct ldlm_ast_work, w_list);
933 /* It's possible to receive a completion AST before we've set
934 * the l_completion_ast pointer: either because the AST arrived
935 * before the reply, or simply because there's a small race
936 * window between receiving the reply and finishing the local
939 * This can't happen with the blocking_ast, however, because we
940 * will never call the local blocking_ast until we drop our
941 * reader/writer reference, which we won't do until we get the
942 * reply and finish enqueueing. */
943 LASSERT(w->w_lock != NULL);
945 LASSERT(w->w_lock->l_blocking_ast != NULL);
946 rc = w->w_lock->l_blocking_ast
947 (w->w_lock, &w->w_desc, w->w_data,
949 } else if (w->w_lock->l_completion_ast != NULL) {
950 LASSERT(w->w_lock->l_completion_ast != NULL);
951 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
959 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
960 "disconnect client\n");
961 LDLM_LOCK_PUT(w->w_lock);
962 list_del(&w->w_list);
963 OBD_FREE(w, sizeof(*w));
968 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
970 ldlm_reprocess_all(res);
971 return LDLM_ITER_CONTINUE;
974 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
978 l_lock(&ns->ns_lock);
979 for (i = 0; i < RES_HASH_SIZE; i++) {
980 struct list_head *tmp, *next;
981 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
982 struct ldlm_resource *res =
983 list_entry(tmp, struct ldlm_resource, lr_hash);
985 ldlm_resource_getref(res);
986 l_unlock(&ns->ns_lock);
987 rc = reprocess_one_queue(res, NULL);
988 l_lock(&ns->ns_lock);
990 ldlm_resource_putref(res);
991 if (rc == LDLM_ITER_STOP)
996 l_unlock(&ns->ns_lock);
1000 void ldlm_reprocess_all(struct ldlm_resource *res)
1002 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1006 /* Local lock trees don't get reprocessed. */
1007 if (res->lr_namespace->ns_client) {
1013 l_lock(&res->lr_namespace->ns_lock);
1014 res->lr_tmp = &rpc_list;
1016 rc = ldlm_reprocess_queue(res, &res->lr_converting);
1017 if (rc == LDLM_ITER_CONTINUE)
1018 ldlm_reprocess_queue(res, &res->lr_waiting);
1021 l_unlock(&res->lr_namespace->ns_lock);
1023 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
1024 if (rc == -ERESTART) {
1025 LASSERT(list_empty(&rpc_list));
1031 void ldlm_cancel_callback(struct ldlm_lock *lock)
1033 l_lock(&lock->l_resource->lr_namespace->ns_lock);
1034 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1035 lock->l_flags |= LDLM_FL_CANCEL;
1036 if (lock->l_blocking_ast) {
1037 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1038 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1039 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1043 LDLM_DEBUG(lock, "no blocking ast");
1046 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1049 void ldlm_lock_cancel(struct ldlm_lock *lock)
1051 struct ldlm_resource *res;
1052 struct ldlm_namespace *ns;
1055 res = lock->l_resource;
1056 ns = res->lr_namespace;
1058 l_lock(&ns->ns_lock);
1059 ldlm_del_waiting_lock(lock);
1061 /* Please do not, no matter how tempting, remove this LBUG without
1062 * talking to me first. -phik */
1063 if (lock->l_readers || lock->l_writers) {
1064 LDLM_ERROR(lock, "lock still has references");
1068 ldlm_cancel_callback(lock);
1070 ldlm_resource_unlink_lock(lock);
1071 ldlm_lock_destroy(lock);
1072 l_unlock(&ns->ns_lock);
1076 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1078 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1084 lock->l_ast_data = data;
1085 LDLM_LOCK_PUT(lock);
1089 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1091 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1092 struct ldlm_lock *lock;
1093 struct ldlm_resource *res;
1095 l_lock(&ns->ns_lock);
1096 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1097 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1098 struct ldlm_lock, l_export_chain);
1099 res = ldlm_resource_getref(lock->l_resource);
1100 LDLM_DEBUG(lock, "export %p", exp);
1101 ldlm_lock_cancel(lock);
1102 l_unlock(&ns->ns_lock);
1103 ldlm_reprocess_all(res);
1104 ldlm_resource_putref(res);
1105 l_lock(&ns->ns_lock);
1107 l_unlock(&ns->ns_lock);
1110 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1113 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1114 struct ldlm_resource *res;
1115 struct ldlm_namespace *ns;
1121 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1122 *flags |= LDLM_FL_BLOCK_GRANTED;
1123 RETURN(lock->l_resource);
1126 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1127 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1129 res = lock->l_resource;
1130 ns = res->lr_namespace;
1132 l_lock(&ns->ns_lock);
1134 old_mode = lock->l_req_mode;
1135 lock->l_req_mode = new_mode;
1136 ldlm_resource_unlink_lock(lock);
1138 /* If this is a local resource, put it on the appropriate list. */
1139 if (res->lr_namespace->ns_client) {
1140 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1141 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1143 /* This should never happen, because of the way the
1144 * server handles conversions. */
1145 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1149 res->lr_tmp = &rpc_list;
1150 ldlm_grant_lock(lock, NULL, 0, 0);
1153 /* FIXME: completion handling not with ns_lock held ! */
1154 if (lock->l_completion_ast)
1155 lock->l_completion_ast(lock, 0, NULL);
1159 ldlm_processing_policy policy;
1160 policy = ldlm_processing_policy_table[res->lr_type];
1161 res->lr_tmp = &rpc_list;
1162 rc = policy(lock, &pflags, 0, &err);
1164 if (rc == LDLM_ITER_STOP) {
1165 lock->l_req_mode = old_mode;
1166 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1169 *flags |= LDLM_FL_BLOCK_GRANTED;
1174 l_unlock(&ns->ns_lock);
1177 ldlm_run_ast_work(ns, &rpc_list);
1181 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1183 char str[PTL_NALFMT_SIZE];
1184 struct obd_device *obd = NULL;
1186 if (!((portal_debug | D_ERROR) & level))
1190 CDEBUG(level, " NULL LDLM lock\n");
1194 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1195 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1197 if (lock->l_conn_export != NULL)
1198 obd = lock->l_conn_export->exp_obd;
1199 if (lock->l_export && lock->l_export->exp_connection) {
1200 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1201 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1202 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1203 lock->l_remote_handle.cookie);
1204 } else if (obd == NULL) {
1205 CDEBUG(level, " Node: local\n");
1207 struct obd_import *imp = obd->u.cli.cl_import;
1208 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1209 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1210 imp->imp_connection->c_peer.peer_ni->pni_name,
1211 lock->l_remote_handle.cookie);
1213 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1214 lock->l_resource->lr_name.name[0],
1215 lock->l_resource->lr_name.name[1]);
1216 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1217 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1218 ldlm_lockname[lock->l_granted_mode],
1219 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1220 if (lock->l_resource->lr_type == LDLM_EXTENT)
1221 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1222 " (req "LPU64"-"LPU64")\n",
1223 lock->l_policy_data.l_extent.start,
1224 lock->l_policy_data.l_extent.end,
1225 lock->l_req_extent.start, lock->l_req_extent.end);
1226 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1227 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1228 lock->l_policy_data.l_flock.pid,
1229 lock->l_policy_data.l_flock.start,
1230 lock->l_policy_data.l_flock.end);
1231 else if (lock->l_resource->lr_type == LDLM_IBITS)
1232 CDEBUG(level, " Bits: "LPX64"\n",
1233 lock->l_policy_data.l_inodebits.bits);
1236 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1238 struct ldlm_lock *lock;
1240 lock = ldlm_handle2lock(lockh);
1244 ldlm_lock_dump(D_OTHER, lock, 0);
1246 LDLM_LOCK_PUT(lock);